Morris analysis with previous model results

Hi all,

I have developed a model with six input parameters and several outputs. The coefficients calculation of the PCE is working well (I am using the quadrature method).

Additionally, to the PCE analysis, I would like to perform a sensitivity analysis using the Morris method. However, the computational cost of the model is high, so I want to use the previous results instead of running several simulations again (one simulation takes around 1 minute using 50 processes). From the manual I noticed that several sensitivity methods can use previous results for their analyses, however Morris method is not one of them (at least as is specified in the manual).

Do you know if there is a way to use my previous results for the Morris analysis? As I am using the Smolyak’s method the trajectories are already defined.

Thanks in advance.

Dear Sebastian

Welcome on UQWorld!

Indeed, once you have built a surrogate model in UQLab, it behaves like any other model. By default the last created MODEL is the one used in the subsequent analysis. So if you define a Morris sensitivity analysis right after building a PCE, it will use this PCE to carry out the analysis.

You can access the list of currently available MODELs using the command uq_listModels within your UQLab session. The one with a tick is the active one.

Now, once you have built a PCE with any method, you should better use Sobol’ indices to perform sensitivity analysis. First because it is much more accurate that the information you could get from Morris method (especially in the case of non linear models). Second, because it comes for free when a PCE is available, as explained in the UQLab sensitivity user manual page 16. Indeed there is a closed-form solution for these sensitivity indices from the polynomial chaos coefficients.

I advise that you have a look at the usage of Sobol’ indices and apply it to your problem (and also Morris method, so that you can compare!). Note that UQLab will automatically handle your multiple outputs (i.e it will build a PCE and compute Sobol’ indices for each output separately).

Finally, if your computational model is rather expensive to run (~2 minutes) you may want to use other PCE construction methods, such as LAR: this way, you can exactly choose the number of runs you want to carry out, as opposed to the quadrature method for which the integration points are fixed. For a 6-dimensional problem, typically with 100 points you may get excellent results already. You can check the quality of the various PCE by looking at the field myPCE.Error in your PCE result.

I hope this information will be useful.
Best regards

Dear Bruno,

Thank you so much for your answer, it’s more than that I expected. I will read carefully the Sobol’ indices, and also I will try with the LAR method.